Poor performance with oneAPI vs. Intel classic compilers

[MHlav]

We are observing very poor performance in our floating point intensive commercial engineering application after upgrading from the latest classic compiler version (19.2) to oneAPI 2024 for both x86 and x64 platforms in the Windows environment.  As a matter of fact, the performance is much worse than code generated by Visual C++ 2022.  In the past, the classic compiler application slightly outperformed the Microsoft code, but the 2024 compiler code is much slower.  In fact, it is on the order of clang 18 in performance (both much slower than VC++).   The application only uses the native processor (no GPU).

Anyone else observing this or have ideas how to solve it?  We are contemplating dropping the Intel compiler because the builds are much slower than VC++ and the application is much slower as well.  With the classic compiler, we could take the slower build times because the application performance was better or similar.

[Mysticial]

Yes, it is known that ICX can underperform ICC - often by large margins. I've had some pretty bad experiences as well that are still ongoing.

But from my experience, if you're willing to take the time to isolate some examples and submit them to Intel they may be willing to diagnose the problem and fix it in a future version of ICX.

[MHlav]

The application is very large.  That will take considerable effort.  Not certain it's worth the effort when VC++ is starting to give comparable performance out of the box without this effort.  To me, this compiler is of beta quality at this point if they are selling it as generating the best Intel code of all compilers.  We waited as long as we could to make the conversion from the classic compiler because we were afraid this may happen as we have had a long history with Intel compilers.